GreenLeaf Law Embed Tiny: Compact Model for Legal Retrieval.

This paper introduces GreenLeaf Law Embed Tiny, a compact embedding model with only 0.6 billion parameters, specifically optimized for information retrieval within the legal domain. Despite its small size, the model demonstrates strong performance, achieving 75.11% on the Massive Legal Embedding Benchmark (MLEB) and 64.38% on MTEB(Law, v1), positioning it competitively among models under 1 billion parameters. The development of GreenLeaf-Tiny involved a sophisticated two-stage training pipeline. Initially, knowledge was distilled from a larger teacher model into the more compact student architecture. This was followed by domain-specific fine-tuning, which incorporated hard negative mining to refine its understanding of legal nuances. A crucial component of its success is a meticulously curated dataset comprising 3.4 million query-passage pairs, including 150,000 human-curated samples spanning diverse legal jurisdictions. Furthermore, the model is designed for efficient inference, supporting multiple quantization levels (BF16, INT8, binary). This architectural choice enables its deployment in environments with limited computational resources, making advanced legal retrieval more accessible. The research provides a detailed analysis of its training methodology, architectural decisions, and comprehensive evaluation across various legal retrieval tasks.